Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.25
1.25 ! oster 1: /* $NetBSD: rf_diskqueue.c,v 1.24 2003/12/29 03:33:47 oster Exp $ */
1.1 oster 2: /*
3: * Copyright (c) 1995 Carnegie-Mellon University.
4: * All rights reserved.
5: *
6: * Author: Mark Holland
7: *
8: * Permission to use, copy, modify and distribute this software and
9: * its documentation is hereby granted, provided that both the copyright
10: * notice and this permission notice appear in all copies of the
11: * software, derivative works or modified versions, and any portions
12: * thereof, and that both notices appear in supporting documentation.
13: *
14: * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
15: * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
16: * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
17: *
18: * Carnegie Mellon requests users of this software to return to
19: *
20: * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
21: * School of Computer Science
22: * Carnegie Mellon University
23: * Pittsburgh PA 15213-3890
24: *
25: * any improvements or extensions that they make and grant Carnegie the
26: * rights to redistribute these changes.
27: */
28:
1.13 oster 29: /****************************************************************************
1.1 oster 30: *
31: * rf_diskqueue.c -- higher-level disk queue code
32: *
33: * the routines here are a generic wrapper around the actual queueing
1.6 oster 34: * routines. The code here implements thread scheduling, synchronization,
1.1 oster 35: * and locking ops (see below) on top of the lower-level queueing code.
36: *
1.13 oster 37: * to support atomic RMW, we implement "locking operations". When a
38: * locking op is dispatched to the lower levels of the driver, the
39: * queue is locked, and no further I/Os are dispatched until the queue
40: * receives & completes a corresponding "unlocking operation". This
41: * code relies on the higher layers to guarantee that a locking op
42: * will always be eventually followed by an unlocking op. The model
43: * is that the higher layers are structured so locking and unlocking
44: * ops occur in pairs, i.e. an unlocking op cannot be generated until
45: * after a locking op reports completion. There is no good way to
46: * check to see that an unlocking op "corresponds" to the op that
47: * currently has the queue locked, so we make no such attempt. Since
48: * by definition there can be only one locking op outstanding on a
49: * disk, this should not be a problem.
50: *
51: * In the kernel, we allow multiple I/Os to be concurrently dispatched
52: * to the disk driver. In order to support locking ops in this
53: * environment, when we decide to do a locking op, we stop dispatching
54: * new I/Os and wait until all dispatched I/Os have completed before
55: * dispatching the locking op.
56: *
57: * Unfortunately, the code is different in the 3 different operating
58: * states (user level, kernel, simulator). In the kernel, I/O is
59: * non-blocking, and we have no disk threads to dispatch for us.
60: * Therefore, we have to dispatch new I/Os to the scsi driver at the
61: * time of enqueue, and also at the time of completion. At user
62: * level, I/O is blocking, and so only the disk threads may dispatch
63: * I/Os. Thus at user level, all we can do at enqueue time is enqueue
64: * and wake up the disk thread to do the dispatch.
1.1 oster 65: *
1.13 oster 66: ****************************************************************************/
1.15 lukem 67:
68: #include <sys/cdefs.h>
1.25 ! oster 69: __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.24 2003/12/29 03:33:47 oster Exp $");
1.1 oster 70:
1.14 oster 71: #include <dev/raidframe/raidframevar.h>
72:
1.1 oster 73: #include "rf_threadstuff.h"
74: #include "rf_raid.h"
75: #include "rf_diskqueue.h"
76: #include "rf_alloclist.h"
77: #include "rf_acctrace.h"
78: #include "rf_etimer.h"
79: #include "rf_general.h"
80: #include "rf_debugprint.h"
81: #include "rf_shutdown.h"
82: #include "rf_cvscan.h"
83: #include "rf_sstf.h"
84: #include "rf_fifo.h"
1.11 oster 85: #include "rf_kintf.h"
1.1 oster 86:
87: static int init_dqd(RF_DiskQueueData_t *);
88: static void clean_dqd(RF_DiskQueueData_t *);
89: static void rf_ShutdownDiskQueueSystem(void *);
90:
1.21 oster 91: #ifndef RF_DEBUG_DISKQUEUE
92: #define RF_DEBUG_DISKQUEUE 0
93: #endif
94:
95: #if RF_DEBUG_DISKQUEUE
1.1 oster 96: #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
97: #define Dprintf2(s,a,b) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),NULL,NULL,NULL,NULL,NULL,NULL)
98: #define Dprintf3(s,a,b,c) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),NULL,NULL,NULL,NULL,NULL)
1.21 oster 99: #else
100: #define Dprintf1(s,a)
101: #define Dprintf2(s,a,b)
102: #define Dprintf3(s,a,b,c)
103: #endif
1.1 oster 104:
1.13 oster 105: /*****************************************************************************
1.1 oster 106: *
1.13 oster 107: * the disk queue switch defines all the functions used in the
108: * different queueing disciplines queue ID, init routine, enqueue
109: * routine, dequeue routine
1.1 oster 110: *
1.13 oster 111: ****************************************************************************/
1.1 oster 112:
1.22 jdolecek 113: static const RF_DiskQueueSW_t diskqueuesw[] = {
1.6 oster 114: {"fifo", /* FIFO */
115: rf_FifoCreate,
116: rf_FifoEnqueue,
117: rf_FifoDequeue,
118: rf_FifoPeek,
1.1 oster 119: rf_FifoPromote},
120:
1.6 oster 121: {"cvscan", /* cvscan */
122: rf_CvscanCreate,
123: rf_CvscanEnqueue,
124: rf_CvscanDequeue,
125: rf_CvscanPeek,
126: rf_CvscanPromote},
127:
128: {"sstf", /* shortest seek time first */
129: rf_SstfCreate,
130: rf_SstfEnqueue,
131: rf_SstfDequeue,
132: rf_SstfPeek,
1.1 oster 133: rf_SstfPromote},
134:
1.6 oster 135: {"scan", /* SCAN (two-way elevator) */
136: rf_ScanCreate,
137: rf_SstfEnqueue,
138: rf_ScanDequeue,
139: rf_ScanPeek,
1.1 oster 140: rf_SstfPromote},
141:
1.6 oster 142: {"cscan", /* CSCAN (one-way elevator) */
143: rf_CscanCreate,
144: rf_SstfEnqueue,
145: rf_CscanDequeue,
146: rf_CscanPeek,
1.1 oster 147: rf_SstfPromote},
148:
149: };
150: #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t))
151:
1.24 oster 152: static struct pool rf_dqd_pool;
1.1 oster 153: #define RF_MAX_FREE_DQD 256
154: #define RF_DQD_INC 16
155: #define RF_DQD_INITIAL 64
156:
157: #include <sys/buf.h>
158:
1.6 oster 159: static int
160: init_dqd(dqd)
161: RF_DiskQueueData_t *dqd;
162: {
1.13 oster 163:
1.9 oster 164: dqd->bp = (struct buf *) malloc(sizeof(struct buf),
165: M_RAIDFRAME, M_NOWAIT);
1.1 oster 166: if (dqd->bp == NULL) {
1.6 oster 167: return (ENOMEM);
1.1 oster 168: }
1.6 oster 169: memset(dqd->bp, 0, sizeof(struct buf)); /* if you don't do it, nobody
170: * else will.. */
171: return (0);
1.1 oster 172: }
173:
1.6 oster 174: static void
175: clean_dqd(dqd)
176: RF_DiskQueueData_t *dqd;
1.1 oster 177: {
1.6 oster 178: free(dqd->bp, M_RAIDFRAME);
179: }
180: /* configures a single disk queue */
1.9 oster 181:
1.7 oster 182: int
1.9 oster 183: rf_ConfigureDiskQueue(
184: RF_Raid_t * raidPtr,
185: RF_DiskQueue_t * diskqueue,
186: RF_RowCol_t c,
1.22 jdolecek 187: const RF_DiskQueueSW_t * p,
1.9 oster 188: RF_SectorCount_t sectPerDisk,
189: dev_t dev,
190: int maxOutstanding,
191: RF_ShutdownList_t ** listp,
192: RF_AllocListElem_t * clList)
1.6 oster 193: {
194: int rc;
195:
196: diskqueue->col = c;
197: diskqueue->qPtr = p;
198: diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp);
199: diskqueue->dev = dev;
200: diskqueue->numOutstanding = 0;
201: diskqueue->queueLength = 0;
202: diskqueue->maxOutstanding = maxOutstanding;
203: diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
204: diskqueue->nextLockingOp = NULL;
205: diskqueue->numWaiting = 0;
206: diskqueue->flags = 0;
207: diskqueue->raidPtr = raidPtr;
1.23 oster 208: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c];
1.25 ! oster 209: rf_mutex_init(&diskqueue->mutex);
1.6 oster 210: rc = rf_create_managed_cond(listp, &diskqueue->cond);
211: if (rc) {
1.18 oster 212: rf_print_unable_to_init_cond(__FILE__, __LINE__, rc);
1.6 oster 213: return (rc);
214: }
215: return (0);
1.1 oster 216: }
217:
1.6 oster 218: static void
219: rf_ShutdownDiskQueueSystem(ignored)
220: void *ignored;
221: {
1.24 oster 222: pool_destroy(&rf_dqd_pool);
1.1 oster 223: }
224:
1.6 oster 225: int
226: rf_ConfigureDiskQueueSystem(listp)
227: RF_ShutdownList_t **listp;
228: {
229: int rc;
230:
1.24 oster 231: pool_init(&rf_dqd_pool, sizeof(RF_DiskQueueData_t), 0, 0, 0,
232: "rf_dqd_pl", NULL);
233: pool_sethiwat(&rf_dqd_pool, RF_MAX_FREE_DQD);
234: pool_prime(&rf_dqd_pool, RF_DQD_INITIAL);
235:
1.6 oster 236: rc = rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
237: if (rc) {
1.18 oster 238: rf_print_unable_to_add_shutdown( __FILE__, __LINE__, rc);
1.6 oster 239: rf_ShutdownDiskQueueSystem(NULL);
240: return (rc);
241: }
1.24 oster 242:
1.6 oster 243: return (0);
244: }
245:
246: int
247: rf_ConfigureDiskQueues(
248: RF_ShutdownList_t ** listp,
249: RF_Raid_t * raidPtr,
250: RF_Config_t * cfgPtr)
251: {
1.23 oster 252: RF_DiskQueue_t *diskQueues, *spareQueues;
1.22 jdolecek 253: const RF_DiskQueueSW_t *p;
1.23 oster 254: RF_RowCol_t r,c;
1.6 oster 255: int rc, i;
256:
257: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
258:
259: for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
260: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
261: p = &diskqueuesw[i];
262: break;
263: }
264: }
265: if (p == NULL) {
266: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
267: p = &diskqueuesw[0];
268: }
1.10 oster 269: raidPtr->qType = p;
1.23 oster 270:
1.24 oster 271: RF_MallocAndAdd(diskQueues,
272: (raidPtr->numCol + RF_MAXSPARE) *
1.23 oster 273: sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *),
274: raidPtr->cleanupList);
275: if (diskQueues == NULL)
1.6 oster 276: return (ENOMEM);
277: raidPtr->Queues = diskQueues;
1.23 oster 278:
279: for (c = 0; c < raidPtr->numCol; c++) {
280: rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c],
281: c, p,
282: raidPtr->sectorsPerDisk,
283: raidPtr->Disks[c].dev,
284: cfgPtr->maxOutstandingDiskReqs,
285: listp, raidPtr->cleanupList);
286: if (rc)
287: return (rc);
1.6 oster 288: }
289:
1.23 oster 290: spareQueues = &raidPtr->Queues[raidPtr->numCol];
1.6 oster 291: for (r = 0; r < raidPtr->numSpare; r++) {
1.9 oster 292: rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
1.23 oster 293: raidPtr->numCol + r, p,
294: raidPtr->sectorsPerDisk,
295: raidPtr->Disks[raidPtr->numCol + r].dev,
296: cfgPtr->maxOutstandingDiskReqs, listp,
297: raidPtr->cleanupList);
1.6 oster 298: if (rc)
299: return (rc);
300: }
301: return (0);
302: }
1.1 oster 303: /* Enqueue a disk I/O
304: *
305: * Unfortunately, we have to do things differently in the different
306: * environments (simulator, user-level, kernel).
307: * At user level, all I/O is blocking, so we have 1 or more threads/disk
308: * and the thread that enqueues is different from the thread that dequeues.
309: * In the kernel, I/O is non-blocking and so we'd like to have multiple
310: * I/Os outstanding on the physical disks when possible.
311: *
312: * when any request arrives at a queue, we have two choices:
313: * dispatch it to the lower levels
314: * queue it up
315: *
316: * kernel rules for when to do what:
317: * locking request: queue empty => dispatch and lock queue,
318: * else queue it
319: * unlocking req : always dispatch it
320: * normal req : queue empty => dispatch it & set priority
321: * queue not full & priority is ok => dispatch it
322: * else queue it
323: *
324: * user-level rules:
325: * always enqueue. In the special case of an unlocking op, enqueue
326: * in a special way that will cause the unlocking op to be the next
327: * thing dequeued.
328: *
329: * simulator rules:
330: * Do the same as at user level, with the sleeps and wakeups suppressed.
331: */
1.6 oster 332: void
333: rf_DiskIOEnqueue(queue, req, pri)
334: RF_DiskQueue_t *queue;
335: RF_DiskQueueData_t *req;
336: int pri;
337: {
338: RF_ETIMER_START(req->qtime);
339: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
340: req->priority = pri;
341:
1.21 oster 342: #if RF_DEBUG_DISKQUEUE
1.6 oster 343: if (rf_queueDebug && (req->numSector == 0)) {
344: printf("Warning: Enqueueing zero-sector access\n");
345: }
1.21 oster 346: #endif
1.6 oster 347: /*
348: * kernel
349: */
350: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
351: /* locking request */
352: if (RF_LOCKING_REQ(req)) {
353: if (RF_QUEUE_EMPTY(queue)) {
1.23 oster 354: Dprintf2("Dispatching pri %d locking op to c %d (queue empty)\n", pri, queue->col);
1.6 oster 355: RF_LOCK_QUEUE(queue);
356: rf_DispatchKernelIO(queue, req);
357: } else {
358: queue->queueLength++; /* increment count of number
359: * of requests waiting in this
360: * queue */
1.23 oster 361: Dprintf2("Enqueueing pri %d locking op to c %d (queue not empty)\n", pri, queue->col);
1.6 oster 362: req->queue = (void *) queue;
363: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
364: }
365: }
366: /* unlocking request */
367: else
368: if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock
369: * when this I/O completes */
1.23 oster 370: Dprintf2("Dispatching pri %d unlocking op to c %d\n", pri, queue->col);
1.6 oster 371: RF_ASSERT(RF_QUEUE_LOCKED(queue));
372: rf_DispatchKernelIO(queue, req);
373: }
374: /* normal request */
375: else
376: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 377: Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col);
1.6 oster 378: rf_DispatchKernelIO(queue, req);
379: } else {
380: queue->queueLength++; /* increment count of
381: * number of requests
382: * waiting in this queue */
1.23 oster 383: Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col);
1.6 oster 384: req->queue = (void *) queue;
385: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
386: }
387: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
1.1 oster 388: }
1.6 oster 389:
1.1 oster 390:
391: /* get the next set of I/Os started, kernel version only */
1.6 oster 392: void
393: rf_DiskIOComplete(queue, req, status)
394: RF_DiskQueue_t *queue;
395: RF_DiskQueueData_t *req;
396: int status;
397: {
398: int done = 0;
399:
400: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
401:
402: /* unlock the queue: (1) after an unlocking req completes (2) after a
403: * locking req fails */
404: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
1.23 oster 405: Dprintf1("DiskIOComplete: unlocking queue at c %d\n", queue->col);
1.20 oster 406: RF_ASSERT(RF_QUEUE_LOCKED(queue));
1.6 oster 407: RF_UNLOCK_QUEUE(queue);
408: }
409: queue->numOutstanding--;
410: RF_ASSERT(queue->numOutstanding >= 0);
411:
412: /* dispatch requests to the disk until we find one that we can't. */
413: /* no reason to continue once we've filled up the queue */
414: /* no reason to even start if the queue is locked */
415:
416: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
417: if (queue->nextLockingOp) {
418: req = queue->nextLockingOp;
419: queue->nextLockingOp = NULL;
1.23 oster 420: Dprintf2("DiskIOComplete: a pri %d locking req was pending at c %d\n", req->priority, queue->col);
1.6 oster 421: } else {
422: req = (queue->qPtr->Dequeue) (queue->qHdr);
423: if (req != NULL) {
1.23 oster 424: Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col);
1.6 oster 425: } else {
426: Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
427: }
428: }
429: if (req) {
430: queue->queueLength--; /* decrement count of number
431: * of requests waiting in this
432: * queue */
433: RF_ASSERT(queue->queueLength >= 0);
434: }
435: if (!req)
436: done = 1;
437: else
438: if (RF_LOCKING_REQ(req)) {
439: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
1.23 oster 440: Dprintf2("DiskIOComplete: dispatching pri %d locking req to c %d (queue empty)\n", req->priority, queue->col);
1.6 oster 441: RF_LOCK_QUEUE(queue);
442: rf_DispatchKernelIO(queue, req);
443: done = 1;
444: } else { /* put it aside to wait for
445: * the queue to drain */
1.23 oster 446: Dprintf2("DiskIOComplete: postponing pri %d locking req to c %d\n", req->priority, queue->col);
1.6 oster 447: RF_ASSERT(queue->nextLockingOp == NULL);
448: queue->nextLockingOp = req;
449: done = 1;
450: }
451: } else
452: if (RF_UNLOCKING_REQ(req)) { /* should not happen:
453: * unlocking ops should
454: * not get queued */
455: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for
456: * the future */
1.23 oster 457: Dprintf2("DiskIOComplete: dispatching pri %d unl req to c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->col);
1.6 oster 458: rf_DispatchKernelIO(queue, req);
459: done = 1;
460: } else
461: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 462: Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col);
1.6 oster 463: rf_DispatchKernelIO(queue, req);
464: } else { /* we can't dispatch it,
465: * so just re-enqueue
466: * it. */
467: /* potential trouble here if
468: * disk queues batch reqs */
1.23 oster 469: Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col);
1.6 oster 470: queue->queueLength++;
471: (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
472: done = 1;
473: }
474: }
475:
476: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
1.1 oster 477: }
478: /* promotes accesses tagged with the given parityStripeID from low priority
479: * to normal priority. This promotion is optional, meaning that a queue
480: * need not implement it. If there is no promotion routine associated with
481: * a queue, this routine does nothing and returns -1.
482: */
1.6 oster 483: int
484: rf_DiskIOPromote(queue, parityStripeID, which_ru)
485: RF_DiskQueue_t *queue;
486: RF_StripeNum_t parityStripeID;
487: RF_ReconUnitNum_t which_ru;
488: {
489: int retval;
490:
491: if (!queue->qPtr->Promote)
492: return (-1);
493: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
494: retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
495: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
496: return (retval);
497: }
498:
499: RF_DiskQueueData_t *
500: rf_CreateDiskQueueData(
501: RF_IoType_t typ,
502: RF_SectorNum_t ssect,
503: RF_SectorCount_t nsect,
504: caddr_t buf,
505: RF_StripeNum_t parityStripeID,
506: RF_ReconUnitNum_t which_ru,
507: int (*wakeF) (void *, int),
508: void *arg,
509: RF_DiskQueueData_t * next,
510: RF_AccTraceEntry_t * tracerec,
511: void *raidPtr,
512: RF_DiskQueueDataFlags_t flags,
513: void *kb_proc)
514: {
515: RF_DiskQueueData_t *p;
516:
1.24 oster 517: p = pool_get(&rf_dqd_pool, PR_WAITOK);
518: if (init_dqd(p)) {
519: /* no memory for the buffer!?!? */
520: pool_put(&rf_dqd_pool, p);
521: return(NULL);
522: }
1.6 oster 523:
524: p->sectorOffset = ssect + rf_protectedSectors;
525: p->numSector = nsect;
526: p->type = typ;
527: p->buf = buf;
528: p->parityStripeID = parityStripeID;
529: p->which_ru = which_ru;
530: p->CompleteFunc = wakeF;
531: p->argument = arg;
532: p->next = next;
533: p->tracerec = tracerec;
534: p->priority = RF_IO_NORMAL_PRIORITY;
535: p->raidPtr = raidPtr;
536: p->flags = flags;
537: p->b_proc = kb_proc;
538: return (p);
539: }
540:
541: void
542: rf_FreeDiskQueueData(p)
543: RF_DiskQueueData_t *p;
1.1 oster 544: {
1.24 oster 545: clean_dqd(p);
546: pool_put(&rf_dqd_pool, p);
1.1 oster 547: }
CVSweb <webmaster@jp.NetBSD.org>